Michael Limberger

Show it a picture

One of the coolest things about local AI is vision models. Upload an image and the model can see it, describe it, read text, and answer questions, all on hardware you own. A regular text model only understands words. A vision model understands words and images: a photo, a screenshot, a chart, handwriting, or a diagram.

Practical uses: a screenshot of an error (explain and suggest fixes), a whiteboard from a meeting (transcribe and organize), a chart (describe trends), handwriting (transcribe), a product photo (listing description), or a foreign menu (translate).

We pulled two options. Qwen3-VL (8B or 4B) is the dedicated specialist: OCR in 32 languages, spatial reasoning, charts, and even video frames if you extract them. 8B is the sweet spot. 4B is lighter. Mistral Small 3.2 is a 24B model that does text and vision in one package. It is not as specialized as Qwen3-VL, but you do not have to switch models when you go from chat to an image.

Paperclip, then ask

Select a vision model. Click the attachment icon. PNG, JPG, JPEG, GIF, and WebP all work. A thumbnail appears. Type a question and press Enter. The model receives both the text and the image.

Demo now: press Cmd+Shift+4, select an area, upload it to Qwen3-VL, and ask "What do you see in this image? Describe everything." It should identify text, UI, colors, and layout. On code screenshots it will often name the language and say what the snippet does.

OCR is optical character recognition: reading text from images. Try a handwritten note, a PDF you cannot copy from, a sign, or a receipt. Prompt: "Transcribe all the text in this image exactly as written." Thirty-two languages cover English, French, Chinese, Arabic, Japanese, and more. For charts, ask for trends, key points, and a short summary. Bar, line, pie, and scatter charts all work to some degree: axes, labels, outliers. It is not perfect. It is often surprisingly good.

You can also make a custom model with Vision enabled in Capabilities. Example: Screenshot Analyst, base qwen3-vl, thorough description including exact text, UI, layout, and code language if present. Switch to it when you need screen analysis.

Limits to know: small text in large images gets missed, so crop. Complex overlapping diagrams confuse the model, so simpler is better. The model sometimes hallucinates details, so verify if accuracy matters. Video is not directly supported; extract frames. Resolution matters, so upload the highest quality you have.

Multimodal means multiple input types. A vision model is multimodal because it handles text and images. Audio exists in the wider ecosystem but is less common locally. In OpenWebUI, multimodal practically means vision. If the selected model does not support it, the upload will not do anything useful. Next: when things break.